[ICRTCET-2018] ISSN 2348 8034 Impact Factor- 5.070 GLOBAL ...

[ICRTCET-2018] ISSN 2348 – 8034 Impact Factor- 5.070

(C)Global Journal Of Engineering Science And Researches

78

GLOBAL JOURNAL OF ENGINEERING SCIENCE AND RESEARCHES

EMOTION DETECTION AND ANALYSIS ON SOCIAL MEDIA Bharat Gaind

*1, Varun Syal

2 & Sneha Padgalwar

3

*1,2&3CSE Department, IIT Roorkee, Roorkee, India

ABSTRACT

In this paper, we address the problem of detection, classification and quantification of emotions of text in any form. We consider English text collected from social media like Twitter, which can provide information having utility in a

variety of ways, especially opinion mining. Social media like Twitter and Facebook is full of emotions, feelings and

opinions of people all over the world. However, analyzing and classifying text on the basis of emotions is a big

challenge and can be considered as an advanced form of Sentiment Analysis. This paper proposes a method to

classify text into six different Emotion-Categories: Happiness, Sadness, Fear, Anger, Surprise and Disgust. In our

model, we use two different approaches and combine them to effectively extract these emotions from text. The first

approach is based on Natural Language Processing, and uses several textual features like emoticons, degree words

and negations, Parts Of Speech and other grammatical analysis. The second approach is based on Machine Learning

classification algorithms. We have also successfully devised a method to automate the creation of the training-set

itself, so as to eliminate the need of manual annotation of large datasets. Moreover, we have managed to create a

large bag of emotional words, along with their emotion-intensities. On testing, it is shown that our model provides

significant accuracy in classifying tweets taken from Twitter.

Keywords: Sentiment Analysis, Machine learning, Data mining, Natural Language Processing. .

I. INTRODUCTION

Emotions are described as intense feelings that are directed at something or someone in response to internal or

external events having a particular significance for the individual. And the internet, today, has become a key

medium through which people express their emotions, feelings and opinions. Every event, news or activity around

the world, is shared, discussed, posted and commented on social media, by millions of people. Eg. “The Syria

chemical attacks break my heart!! :’(” or “Delicious dinner at Copper Chimney! :D” or “OMG! That is so scary!”.

Capturing these emotions in text, especially those posted or circulated on social media, can be a source of precious

information, which can be used to study how different people react to different situations and events.

Business analysts can use this information to track feelings and opinions of people with respect to their products.

The problem with most of the Sentiment Analysis that is done today is that the analysis only informs whether the

public reaction is positive or negative but fails to describe the exact feelings of the customers and the intensity of

their reaction. With our emotional analysis, they can have a more profound analysis of their markets than the naive

2-way Sentiment Anal-ysis, which itself has turned their businesses more profitable. Business leaders can analyse

the holistic view of people in response to their actions or events and work accordingly. Also, health-analysts can

study the mood swings of individuals or masses at different times of the day or in response to certain events. It can

also be used to formulate the mental or emotional state of an individual, studying his/her activity over a period of

time, and possibly detect depression risks.

There are plenty of research works that have focussed on Sentiment Analysis and provide a 2-way classification of

text. But few have actually focussed on mining emotions from text. However, machine analysis of text to classify

and score it on the basis of emotions poses the following challenges :

Instead of the usual two categories in Sentiment Analysis, there are six Emotion-Categories in which we need

to classify the tweets.

Lack of manually annotated data to train classifiers to label data into six categories.

Unavailability of a comprehensive bag of Emotion-words labeled and scored according to Emotion-Categories

(Happiness, Sadness, etc.) and their intensities, that can be used to detect Emotion-words in text.



79

In order to address the aforementioned challenges, it was important to devise a system that could generate a good

and reliable training-set for the classifier, a labeled bag of words, and an algorithm that could not only detect emotions, but also score and label the tweets according to those emotions.

A lot of research has been done on classifying comments, opinions, movie/product reviews, ratings,

recommendations and other forms of online expression into positive or negative sentiments. Emotions have also

been studied, but in a limited extent, such as by asking specific questions and judging on the basis of replies, or an

analysis done only on short one-lined headlines or a few others [1], [2], [3], all of which depended on the manual

annotation of the training dataset of a small size and limited scope.

In this paper, we propose a method to classify and quantify tweets according to six standard emotions suggested by

Paul Ekman [4]. Here, we base our analysis on tweets posted on Twitter, but it can be easily extended to any kind of

text whether it is one lined headlines, messages and posts on social media or larger chunks of writings, because of automatic development of our training set. Our main contributions are listed below:

We have developed a system that could score and label any piece of text, especially tweets and posts on

social media according to six Emotion-Categories: Happiness, Sadness, Fear, Surprise, Anger and Disgust

along with their intensity scores, making use of its textual features, a variety of NLP tools and standard

Machine Learning classifiers.

Another significant contribution is that we have success-fully devised a system that could automatically

(without any manual effort) build an efficient training set for our ML Classifiers, consisting of a large

enough set of labeled tweets from all Emotion-Categories.We have created a large bag of words in English,

that consists of words expressing a particular emotion along with the intensity of that emotion.

We were able to achieve an accuracy of about 91.7% and 85.4% using J48 and SMO classifiers

respectively using the training set we built.

II. RELATED WORK

In the recent past, with the rise of social media such as blogs and social networks, a lot of interest has been fueled in

Sentiment Analysis. Lately, a lot of research has been done on classifying comments, opinions, movie/product

reviews, rat-ings, recommendations and other forms of online expressions into positive or negative sentiments.

Earlier research involved manually annotated corpus of limited size to classify the emotions. Wiebe et al. [5] worked

on the manual annotation of emotions, opinions and sentiments in a sentence corpus (of size 10,000) of news

articles. Segundo et al. [6] also studies the presence of emotions in text, and is a functional theory of the language used for expressing attitudes, judgments and emotions [3]. This paper deals explicitly with emotions, which none of

the aforementioned works do. There was, however, one research work [7], that classified text into six Emotion-

Categories, but that was only limited to classification of news headlines, and the training set used was created

manually. We, on the other hand, have developed a system, which classifies text in any form (eg. news, tweets, or

narrative) and uses a training set, which is generated automatically. This saved a lot of effort. Another similar work

was of Carlo and Rada [8] who did a similar classification on news headlines but even they used manually annotated

corpus for their classification. There is one research work [9], that explores the possibility of creating automatic

training datasets, like we do and also tries to find out if creating large emotion datasets can increase the emotion-

detection accuracy in tweets. Some aspects of their work are similar to ours, but the emotion-detection accuracy we

have achieved is much higher. Another similar work is [10], but again, our accuracy is much higher. A recent work

[11] explores the possibility of predicting future stock returns based on tweets related to presidential elections and

NASDAQ-100 companies. Another recent work [12] uses convolutional neu-ral network architecture for emotion identification in Twitter messages. Their approach uses unsupervised learning, whereas we use supervised learning

and their accuracy of 55.77% is much lower than ours.

III. DATA SETS

This section describes the various data sets used, such as Tweets Set, Emotion-Words Set (EWS), Degree-Words Set

and Location-Areas Set.



80

A. Tweets Set

We use Tweepy [13] to collect tweets, which is a Python library for accessing the Twitter API. It takes as input various parameters, such as coordinates, radius, etc., and after removal of duplicates, links, hashtags, and words in

other languages (besides English) from these tweets, stores the tweet-ids, text and location of the most recent ones

in the database. This provided us with a list of tweets from various locations across the country. Eg. The Tweets

Set, we created for Delhi has about 10,000 entries. Another way we used Tweepy is by feeding it a twitter-

username (of a user) as an input to store all the tweets of that user (till date), in our database.

TABLE I: An example of Emotion-Words Set (EWS)

Word/

Emoticon

Emotion-

Category

Intensity-

Category

:O SURPRISE STRONG

Repugnance DISGUST STRONG

Delighted HAPPINESS MEDIUM

Afraid FEAR STRONG

:( SADNESS STRONG

Irritated ANGER STRONG

Lucky HAPPINESS LIGHT

B. Emotion-Words Set (EWS)

A proper selection of relevant and commonly-used Emotion-words is one of the most essential and indispensable

aspects in the emotional quantification of a sentence. We have created a high-quality accurate bag of words, called

Emotion-Words Set (EWS), of around 1500 words. It has been developed by recursively searching (Depth First

Search) the synonyms of the 6 basic Emotion-Categories (HAPPINESS, SADNESS, ANGER, SURPRISE, FEAR,

DISGUST) in a thesaurus [14], up to two levels. Each of these words was then manually labeled to one of the three

Intensity-Categories (STRONG, MEDIUM, LIGHT), as shown in Table I.

C. Degree-Words Set

Degree-Words Set is a set of about 50 degree words, that are used to strengthen or weaken the intensity of emotions in a sentence. Eg. “too happy” and “hardly happy” have two different meanings, almost opposite to

each other. In this set, each word has an associated Degree-Intensity stored with it; H



81

Fig. 1: An overview of the first approach

meaning High (having a high multiplying effect), N meaning Negation (having an opposing effect), and L meaning

Low (having a Low multiplying effect). Examples include words like, “too”, “more” (H); “hardly”(N); and “nearly” (L), etc.

D. Location-Areas Set

We also store the areas of about 20 major cities in India in Location-Areas Set, to calculate the radii that are to be

used during tweet extraction. Examples are New Delhi (1484 sq. km.), Mumbai (603 sq. km.), etc.

IV. EMOTION-DETECTION ALGORITHM

Our model consists of two completely different, yet interde-pendent approaches. The first approach uses Natural Language Processing, Emotion-Words Set and several textual features. It attempts to classify and score text

according to the emotions present in it. The second approach uses standard classifiers like SMO and J48 to classify

tweets. Finally, we combine both these approaches to propose a Hybrid approach to detect emotions in text more

effectively. Note that, even though we are extensively using the tweets example throughout this paper, this algorithm

is very generic and can be used to detect and quantify emotions in any piece of text.

A. Using NLP and EWS: First Approach

The first approach uses the tweets in Tweets Set, which are already free from any unwanted characters, hyperlinks or

hashtags. Various Stanford CoreNLP [15] tools are used for a comprehensive linguistic analysis. The following

steps comprise the first approach.

Tokenizing and Annotating: First, we use PTBTokenizer [16] to tokenize the tweets into sentences, which are further tokenized into tokens. Then, we remove all the stop

words from these tokens. The filtered tokens are then annotated using the following CoreNLP annotators:

pos: Parts of Speech (noun, verb, adjective, for example)

lemma: Lemmatized version of that word.

ner: For Named Entitiy Recognition.

We also fetch the Grammatical dependencies between the tokens in a sentence which include nsubj (the

subject of a sentence), dObj (the object of a sentence), advMod (the relation between a word and an adverb),

and neg (the relation between a word and a negative word).

Finding Hits: In this step, the annotated (and filtered) tokens are matched against the words present in the

EWS. But first, all the words in the EWS are lemmatized to their base form. Eg. “happiness” and “happily”



82

are changed to “happy”. While matching the tokens against the EWS, only the tokens that are annotated as

“O” (the other entity in Named Entity Recognition) are considered, because a named entity (location, time or a person word, for example) can never be an Emotion-word. A matched token along with all its

characteristics/annotations is stored as a hit.

Detecting person in context: In our analysis, we aim at detecting the emotions expressed by the tweet’s poster,

stress-ing on the poster’s feelings. Therefore, we try to differentiate between cases in which the emotions involved

are in relation to the poster himself or someone else. For instance, the degree of sadness of the poster in the tweet ”I

am sad” is much higher compared to the tweet, ”He is sad”, posted by the same person, as the former clearly and

directly suggests the sadness of the poster himself. This is done by using the nsubj annotation of the tokens, which

tells us who the subject of the sentence is. After finding the subject of the sentence, we determine which person

(first, second or third), that subject belongs to. For this, we maintain a list of first (“I”, “me”, etc.), second (“You”,

“your”, etc.) and third person (“He”, “She”, “They”, etc.) pronouns. Also, if the nsubj is a proper noun, then the person of that nsubj is considered to be third. Next, each hit detected in the previous step is associated with a

person, by finding the nearest nsubj to the hit on its left side. If there is no nsubj in a sentence, we associate the hit

with the first person. For instance, the hit “Nice” in the tweet, “Nice to see you!!” is associated to the first person,

as there is no nsubj in the sentence.

Effect of Negation and Degree-words: There are many instances, when some words detected using the EWS are ac-

tually used in an entirely opposite sense, because of negations like “not”, “never”, “don’t”, etc. For instance, the

tweet “Not at all feeling excited for school!” may result in getting Happiness as its Emotion-Category (because of

the Emotion-word “excited”), if negations are not accounted for. Also, there are many words (mostly adverbs), that

enhance the intensity of an emotion. Eg. “I feel good” and “I feel too good” contain the same Emotion-word

“good” but the score given to the second sentence should be more. For detecting negations and degree-words, we

search for tokens that have been annotated as neg and advmod, while determining the grammatical de-pendencies in step 1, and tag them as degree/negation words, if they lie in the Degree-Words Set. The Degree-Intensity of the

degree-word associated (if any) with every hit is stored as an annotation to the hit.

Emoticon Detection: Emoticons are a very useful and informative source for detecting emotions in text. Since emoti-

cons are direct ways of knowing the emotions of the user, they are given a heavy weightage, if found. This feature is

very effi-cient and accurate, especially with the ever-rising popularity of emoticons. We have added a list of 100+

most commonly used emoticons in the EWS to enhance our model, along with their Intensity-Categories (STRONG

and MEDIUM). The emoticon-detection process is done before tokenization and uses regex. The emoticons detected

in a tweet are also treated as hits.

TABLE II: Emotion-Scores of various feature combinations

Intensity-

Category

Degree-

Intensity

emotScor

e

STRONG None 6

STRONG H 8

STRONG L 6

STRONG N 2*

MEDIUM None 4

MEDIUM H 6



83

MEDIUM L 6

MEDIUM N 4*

LIGHT None 2

LIGHT H 6

LIGHT L 4

LIGHT N 4*

6) Scoring: Every hit is finally, a tuple of its features that contribute in the scoring process:

• lemma: the lemmatized version of the hit word

• Emotion-Category: HAPPINESS, SADNESS, ANGER, SURPRISE, FEAR, DISGUST

TABLE III: Emoticon Scores

TABLE IV: Person Scores

• Intensity-Category: STRONG, MEDIUM or LIGHT

• Degree-Intensity: H, L, N or None

• person: first, second or third

Based on the various possible combinations of these fea-tures, emotScore and perScore is calculated for each of

the hits. If the hit is an English word (and not an emoticon), the values in Table II are used for calculating

emotScore, whereas for emoticons, Table III is used. It should be noted that if the Degree-Intensity of a hit is N (negation ef-fect), its Emotion-Category changes. This means that HAPPI-NESS turns into SADNESS/ANGER and

vice-versa (with their emotScores given in Table II, marked by asterisks), whereas the emotScores of the remaining

three Emotion-Categories (DISGUST, SURPRISE, FEAR) becomes zero, when the Degree-Intensity is N. Table IV

is used for calculating the perScore of a hit. The final emotional score (Score) of a tweet is a six-tuple, which is

calculated by summing over all the hits (of a particular Emotion-Category) in all the sentences in the tweet for

each of the six Emotion-Categories (ECj, j=1,2,...6).

All Hits of ECj

Score[ECj] = (emotScore ∗ perScore) (1)

Score[ECj]

(2) RelScore[ECj] = ∗ 100



84

6

Score[ECj] j=1

Also, the relative score (RelScore) is calculated to under-stand the percentage of each emotion in a tweet. If there

are no Emotion-words in a tweet, the value of scores of all Emotion-Categories in Score will be zero. For example,

if the tweet has no words of the Emotion-Category FEAR, then Score[FEAR] is 0. An overview of the first

approach is given in Fig 1.

B. Using Machine Learning: Second Approach

We use another method for emotion-detection where we use a Machine Learning classifier. For this, the most

important and critical step is to prepare a good training-set. We make use of the first approach to generate the

training-set, which is free from any manual annotation and thus can be developed or updated quickly.

1) Generation of the training-set: We have chosen a set of seed words, comprising of commonly used Emotion-

words and emoticons from the EWS, evenly distributed over all the Emotion-Categories. We then query Tweepy

using these seed words to develop a huge database of around 13,000 tweets. The seed words are used to ensure that

we get tweets that express at least one of the six emotions. The retweets are removed to avoid any repetition of

tweets. The even distribution of seed words ensures that we get an even distribution of tweets over all the Emotion-

Categories, so that our classifier isn’t biased. However, we have analysed that there is a large proportion of tweets

on Twitter expressing happiness, so we have kept the number of seed words possessing the Emotion-Category

HAPPINESS, on the higher side.

2) Filtering and Labelling the training-set: For training an accurate classifier, we need all the tweets in the

training-set to be strongly expressing only one of the six Emotion-Categories. However, there are many tweets comprising of more than one emotion. Such type of tweets reduces the accuracy of the classifier, as we can give

only one label to a tweet in the training set. If a tweet contains more than one emotion in the training-set, it will

result in faulty learning as words related to the other (unlabeled) Emotion-Category will also be treated as those

related to the category labeled. Thus, we use the first approach to label all the 13,000 tweets and filter out the tweets

having none or mixed emotions. Only those tweets which have a percentage of more than 70% for a particular

emotion are labeled and fed to the classifier for training. The final distribution of tweets in our labeled training set is

as shown in Table V.

TABLE V: Distribution of tweets in the training-set

Emoticon-

Category

No. of

tweets

HAPPINESS 2617

SADNESS 1416

FEAR 1459

DISGUST 776

ANGER 1316

SURPRISE 944

Total 8528



85

3) Training the classifiers: We use the open source library Weka [17] for the implementation of the classifiers. We use two very popular Weka classifiers, SMO [18] and J48 [19]. Before classification, we applied the following

pre-processing steps on the data:

1. Stop-Word Filtering

2. Lower-casing all words

3. Stemming each word using Weka’s Snowball Stemmer

The classifiers output the relative probabilities of each of the six Emotion-Categories, for a tweet (or a piece

of text). The Emotion-Category with the highest probability is called the Labeled-Category (Lc).

C. Combining First and Second Approach Results

We finally combine the first and the second approach. The scores which are generated by the first approach are modified, according to the Labeled-Category of the classifier. The score of only the Labeled-Category is

modified as follows:

F inalScore[Lc] = Score[Lc] + (0.2 ∗ Score[M c]) (3)

where Mc refers to the label with the maximum score, after the first approach results. This is done to give

weightage to the results of the classifier by increasing the score of the classified category in proportion to the

maximum so as to avoid too little or too much relative change in the scores. This will definitely make a

difference in deciding the final Emotion-Category, when two different Emotion-Categories are too close by or

equal and are in competition. After this, the Emotion-Category with the maximum final score is decided as the

final Emotion-Category of the tweet (or the piece of text). An overview of the combined approach is shown in

Fig 2.

V. RESULTS AND ANALYSIS

A. Testing Results

Similar to the training-set, we have created a testing-set of tweets by extracting tweets using some seed words

and then using the first approach to filter and label these emotional tweets. This set consists of a total of 900

tweets, where each Emotion-Category has around 150 tweets, so as to maintain uniformity. Also, it is ensured

that all tweets in the testing set are different from those in the training set. The two chosen classifiers, on

testing with the testing-set, gave the results as shown in Table VI and VII. The correctly classified instances

are the tweets for which the expected Emotion-Category matches the actual Emotion-Category. As can clearly be seen from the tables, we have achieved a remarkable accuracy of 91.7% for SMO and 85.4% for J48,

which proves the merits of our Emotion-Detection Algorithm.

TABLE VI: Accuracy of the SMO Classifier



86

B. Surety Factor (Sf )

The surety factor indicates how confident and assertive our analysis and results are, on a given text/tweet. Its value is low, when there is a mismatch between the results of the two approaches or when the text seems to

lack any emotions. On

Fig. 2: Combining the first and the second approach

TABLE VII: Accuracy of the J48 Classifier

the other hand, its value is high when the two approaches concur with each other in results, there are too many hits

or when one of the emotion-scores is very high. The surety factor is calculated on a scale of 6 and is dependent on

several factors:

Classifier label match: Whether the classifier label matches the category of the maximum score of the first

approach. (Value: True/False).

Max score: The value of the maximum of all the six scores.

Max percent: The relative percentage of the Max score over all six scores.

Second diff: The difference between the maximum and second maximum score value as a percent of the

maxi-mum.

Hits: The number of Emotion-words matched in the text.



87

Surety factor is calculated differently for the following two cases:

• If the hits in the tweet/text belong to the same Emotion- Category, then only the factors Classifier label

match and Max score are used.

• If the hits in the tweet/text belong to different Emotion-Categories, all five factors are considered.

C. Visualization of Results

In order to demonstrate the usefulness of the proposed Emotion-Detection Algorithm, we have implemented the

fol-lowing applications:

• One-user Analysis (twitter account): We have developed an interface, which takes as input, the username of

any twitter-user, processes all his/her tweets till date and displays a time-varying mood-swings plot as well as the

relative and absolute emotional distribution in the form of pie charts. Fig. 3 is an example of the varying happiness of a twitter-user over time.

• Location Analysis: Using Google Maps API, a world map is displayed and each circle on it represents the

emotional analysis of that particular region. The radius of each circle is proportional to the area of the region. This

analysis is done for around 20 major cities in India, using the Location-Areas Set (See Fig. 4)

• Document Analysis: Another interface takes as input, entire documents or blocks of text, processes it, and

displays the relative and absolute emotional distribution of the document in the form of pie charts. (See Fig. 5 for

an example sentence.)

Fig. 3: Time varying happiness plot of a twitter-user



88

Fig. 4: Location-wise emotion-analysis of major cities in India

Fig. 5: Detecting emotions in a piece of text

VI. CONCLUSION

In this paper, we have addressed the problem of classifying text into the six basic Emotion-Categories, rather

than just la-beling them as positive or negative. Through our research and a self-generated reliable bag of

emotional words (EWS), we can now effectively quantify various emotions in any block of text. We have also automatically generated a labeled training-set (without manually labeling the tweets) of emotionally-biased

tweets using a keyword-matching approach, which was then used to train various classifiers. Moreover, we

have also intro-duced the concept of Surety Factor to suggest the reliability of our output and the degree of

usefulness and correctness of our results. Finally, we visualized our results using pie-charts, bar-graphs and

maps, and demonstrated the various applications of our analysis. In future, a system could be established for

automatically updating the bag-of-words which we created, on the basis of new tweets and data analysed.

Using our approach, many interesting apps can be created, such as an add-on to a social-networking site



89

displaying the recent mood of each of your friends. Also, our analysis of Twitter can be extended to the

development of a real-time system, analyzing mood-swings and emotions on Twitter.

REFERENCES 1. Rakesh C Balabantaray, Mudasir Mohammad, and Nibha Sharma. Multi-class twitter emotion

classification: A new approach. International Journal of Applied Information Systems, 4(1):48–53, 2012.

2. Cecilia Ovesdotter Alm, Dan Roth, and Richard Sproat. Emotions from text: machine learning for text-

based emotion prediction. In Proceedings of the conference on human language technology and empirical

methods in natural language processing, pages 579–586. Association for Computational Linguistics, 2005.

3. Lisa Pearl and Mark Steyvers. Identifying emotions, intentions, and attitudes in text using a game with a

purpose. In Proceedings of the naacl hlt 2010 workshop on computational approaches to analysis and generation of emotion in text, pages 71–79. Association for Computa-tional Linguistics, 2010.

4. Paul Ekman. An argument for basic emotions. Cognition & emotion, 6(3-4):169–200, 1992.

5. Janyce Wiebe, Theresa Wilson, and Claire Cardie. Annotating expres-sions of opinions and emotions in

language. Language resources and evaluation, 39(2-3):165–210, 2005.

6. Paulo Roberto Gonc¸alves Segundo. The language of evaluation: Ap-´

7. praisal in english (martin, jr & white, rrr). Linha D’Agua, (21):133–137, 2008.

8. Carlo Strapparava and Rada Mihalcea. Learning to identify emotions in text. In Proceedings of the 2008

ACM symposium on Applied computing, pages 1556–1560. ACM, 2008.

9. Carlo Strapparava and Rada Mihalcea. Semeval-2007 task 14: Affective text. In Proceedings of the 4th

International Workshop on Semantic Evaluations, pages 70–74. Association for Computational

Linguistics, 2007. 10. Wenbo Wang, Lu Chen, Krishnaprasad Thirunarayan, and Amit P Sheth. Harnessing twitter” big data”

for automatic emotion identification. In Privacy, Security, Risk and Trust (PASSAT), 2012 International

Con-ference on and 2012 International Confernece on Social Computing (SocialCom), pages 587–592.

IEEE, 2012.

11. Saima Aman and Stan Szpakowicz. Identifying expressions of emotion in text. In International Conference

on Text, Speech and Dialogue, pages 196–205. Springer, 2007.

12. Clare H Liu. Applications of twitter emotion detection for stock market prediction. PhD thesis,

Massachusetts Institute of Technology, 2017.

13. Dario Stojanovski, Gjorgji Strezoski, Gjorgji Madjarov, and Ivica Dim-itrovski. Emotion identification in

twitter messages for smart city applications. In Information Innovation Technology in Smart Cities, pages

139–150. Springer, 2018.

14. ”Tweepy”. http://www.tweepy.org/. 15. ”Thesaurus.com”. http://www.thesaurus.com/.

16. Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David

McClosky. The Stanford CoreNLP natural language processing toolkit. In Association for Computational

Linguistics (ACL) System Demonstrations, pages 55–60, 2014.

17. ”Stanford Tokenizer”. https://nlp.stanford.edu/software/tokenizer.shtml.

18. Ian H Witten, Eibe Frank, Mark A Hall, and Christopher J Pal. Data Mining: Practical machine learning

tools and techniques. Morgan Kaufmann, 2016.

19. ”Sequential minimal optimization”. https://en.wikipedia.org/wiki/ sequential minimal optimization.

20. ”J48”. http://weka.sourceforge.net/doc.dev/weka/classifiers/trees/ j48.html.

.

Date post:	18-Dec-2021
Category:	Documents
Upload:	others
View:	9 times
Download:	0 times

[ICRTCET-2018] ISSN 2348 8034 Impact Factor- 5.070 GLOBAL ...

Documents